Method and apparatus for estimating a distribution of response times of a storage system for a proposed workload
A distribution of response times of a storage system can be estimated for a proposed workload using a trained learning process. Collections of information about operational characteristics of multiple storage systems are obtained, in which each collection includes parameters describing the configuration of the storage system that was used to create the collection, workload characteristics describing features of the workload that the storage system processed, and storage system response times. For each collection, workload characteristics are aggregated, and the storage system response information is used to train a probabilistic mixture model. The aggregated workload information, storage system characteristics, and probabilistic mixture model parameters of the collections form training examples that are used to train the learning process. Once trained, the learning process is used to provide a distribution of response times that would be expected from a storage system having a proposed configuration when processing a proposed workload.
1. A non-transitory tangible computer readable storage medium having stored thereon a computer program for estimating a distribution of response times of a storage system for a proposed workload, the computer program including a set of instructions which, when executed by a computer, cause the computer to perform a method comprising the steps of:
obtaining a large number of collections of storage system response data from a plurality of storage systems during respective collection periods, each of the collections of storage system response data including: information describing physical hardware configuration parameters of a respective storage system that was used to create the collection and a plurality of workload snapshots obtained from the respective storage system, each workload snapshot including a number of read operations and a number of write operations on the storage system during a time interval within the respective collection period, and an average storage system response time of the storage system to the workload during the time interval;
for each collection, aggregating the number of read operations during the collection period and aggregating the number of write operations on the storage system from the plurality of workload snapshots obtained from the respective storage system during the collection period to create aggregated workload feature information for the collection;
for each collection, using the average storage system response times of the plurality of workload snapshots to learn mean and standard deviation coefficients describing two or more distributions of a probability mixture model characterizing storage system response times over the collection period for the collection, the probability mixture model being a probabilistic model describing the storage system response times based on a mixture of the two or more distributions;
creating a respective training example from each collection, each respective training example correlating storage system physical hardware configuration parameters and aggregated workload feature information with the mean and standard deviation coefficients describing the two or more distributions of the probability mixture model characterizing the storage system response times;
using the training examples to train a learning process to cause the learning process to learn a regression between storage system physical hardware configuration parameters and aggregated workload feature information with the mean and standard deviation coefficients describing the two or more distributions of the probability mixture model characterizing the storage system response times; and
using the trained learning process to generate a distribution of expected storage system response times for a storage system having proposed physical hardware configuration parameters and a proposed workload.
2. The non-transitory tangible computer readable storage medium of claim 1 , wherein each collection of storage system response data includes data obtained from a different storage system.
3. The non-transitory tangible computer readable storage medium of claim 1 , wherein each collection of storage system response data includes data from a particular storage system, wherein each collection period is a several hour period of time, and wherein the time intervals of a given collection are several minute time windows of equal size.
4. The non-transitory tangible computer readable storage medium of claim 1 , wherein each snapshot includes an aggregation of a number of operations of a set of additional workload features during the time interval.
5. The non-transitory tangible computer readable storage medium of claim 1 , wherein using the average storage system response times of the plurality of workload snapshots to learn mean and standard deviation coefficients describing two or more distributions of a probability mixture model comprises iteratively using the average storage system response times from the set of snapshots of the collection to converge on the mean and standard deviation coefficients describing the two or more distributions of the probability mixture model for the collection.
6. The non-transitory tangible computer readable storage medium of claim 1 , further comprising testing the learning function using a set of testing examples prior to using the trained learning process.
7. The non-transitory tangible computer readable storage medium of claim 1 , wherein using the trained learning process comprises:
providing, as input to the trained learning process, the proposed storage system configuration and proposed workload; and
comparing the distribution of expected storage system response times generated by the trained learning process to a required distribution of storage system response times.
8. The non-transitory tangible computer readable storage medium of claim 7 , if the expected storage system response times generated by the trained learning process does not sufficiently correlate with the required distribution of expected storage system response times, the method further comprising:
automatically adjusting the proposed storage system configuration to create an adjusted storage system configuration;
providing, as input to the trained learning process, the adjusted storage system configuration and proposed workload;
generating, by the trained learning process, an adjusted distribution of expected storage system response times for the adjusted storage system configuration and proposed workload;
comparing the adjusted distribution of expected storage system response times generated by the trained learning process to the required distribution of expected storage system response times; and
iterating the steps of automatically adjusting, providing, generating, and comparing until the adjusted storage system configuration sufficiently correlates with the required distribution of storage system response times.
9. A method of estimating a distribution of response times of a storage system for a proposed workload, the method comprising:
obtaining a large number of collections of storage system response data from a plurality of storage systems during respective collection periods, each of the collections of storage system response data including: information describing physical hardware configuration parameters of a respective storage system that was used to create the collection and a plurality of workload snapshots obtained from the respective storage system, each workload snapshot including a number of read operations and a number of write operations on the storage system during a time interval within the respective collection period, and an average storage system response time of the storage system to the workload during the time interval;
for each collection, aggregating the number of read operations during the collection period and aggregating the number of write operations on the storage system from the plurality of workload snapshots obtained from the respective storage system during the collection period to create aggregated workload feature information for the collection;
for each collection, using the average storage system response times of the plurality of workload snapshots to learn mean and standard deviation coefficients describing two or more distributions of a probability mixture model characterizing storage system response times over the collection period for the collection, the probability mixture model being a probabilistic model describing the storage system response times based on a mixture of the two or more distributions;
creating a respective training example from each collection, each respective training example correlating storage system physical hardware configuration parameters and aggregated workload feature information with the mean and standard deviation coefficients describing a the two or more distributions of the probability mixture model characterizing the storage system response times;
using the training examples to train a learning process to cause the learning process to learn a regression between storage system physical hardware configuration parameters and aggregated workload feature information with the mean and standard deviation coefficients describing the two or more distributions of the probability mixture model characterizing the storage system response times; and
using the trained learning process to generate a distribution of expected storage system response times for a storage system having proposed physical hardware configuration parameters and a proposed workload.
10. The method of claim 9 , wherein each collection of storage system response data includes data obtained from a different storage system.
11. The method of claim 9 , wherein each collection of storage system response data includes data from a particular storage system, wherein each collection period is a several hour period of time, and wherein the time intervals of a given collection are time windows of equal size.
12. The method of claim 9 , wherein each snapshot includes an aggregation of a number of operations of a set of additional workload features during the time interval.
13. The method of claim 9 , wherein using the average storage system response times of the plurality of workload snapshots to learn mean and standard deviation coefficients describing two or more distributions of a probability mixture model comprises iteratively using the average storage system response times from the set of snapshots of the collection to converge on the mean and standard deviation coefficients describing the two or more distributions of the probability mixture model for the collection.
14. The method of claim 9 , further comprising testing the learning function using a set of testing examples prior to using the trained learning process.
15. The method of claim 9 , wherein using the trained learning process comprises:
providing, as input to the trained learning process, the proposed storage system configuration and proposed workload; and
comparing the distribution of expected storage system response times generated by the trained learning process to a required distribution of storage system response times.
16. The method of claim 15 , if the expected storage system response times generated by the trained learning process does not sufficiently correlate with the required distribution of expected storage system response times, the method further comprising:
automatically adjusting the proposed storage system configuration to create an adjusted storage system configuration;
providing, as input to the trained learning process, the adjusted storage system configuration and proposed workload;
generating, by the trained learning process, an adjusted distribution of expected storage system response times for the adjusted storage system configuration and proposed workload;
comparing the adjusted distribution of expected storage system response times generated by the trained learning process to the required distribution of expected storage system response times; and
iterating the steps of automatically adjusting, providing, generating, and comparing until the adjusted storage system configuration sufficiently correlates with the required distribution of storage system response times.